Distillation

Date: Sep 16, 2026

• 🟡 (13:20) A search for headroom found exactly 25 matches: packages/core/test/cgroup-memory.test.ts:275 tests v1 headroom and treats an oversized unlimited sentinel as max; packages/core/test/embedding-pool.test.ts:792 tests staying at one worker when current cgroup headroom cannot fit a second; :815 tests reserving headroom while an admitted worker is still starting; :829 explains that the same headroom must not be spent again after initialization; :1601 sets _setPoolFreememForTest(0) so no headroom yields ceiling 1; packages/core/test/embedding-pool-memory.test.ts:149 tests the token floor when constrained primary-worker headroom is unknown.
• 🟡 (13:20) Remaining headroom matches in packages/core/test/embedding-cap.test.ts: line 56 tests clamping to current cgroup headroom rather than total limit; line 61 tests discovered headroom when the hard limit is unlimited; line 82 tests using the tighter Linux or Node headroom while ignoring Bun upstream; line 92 tests libuv headroom on constrained non-Linux Node runtimes; line 101 tests not inventing headroom on an unconstrained host; line 249 documents × 0.85 headroom minus baseline over K, below MODEL_MAX_TOKENS; line 397 states that just under two budgets still permits only 1 because a full budget of headroom is required.
• 🟡 (13:20) Remaining headroom matches outside embedding memory tests: packages/core/test/distillation.test.ts:255 sets a 2s threshold as ample jitter headroom over a healthy <1ms; packages/core/test/gradient.test.ts has matches at lines 567, 618, 624, 2077, 2445, 4823, 4914, 4917, 5264, 5773, and 5780 concerning boundary/growth headroom, layer budgets, cache-stable pins, eviction headroom, passthrough behavior, and a calibrate(50_000, SID, msgs.length) case under a 200K l0cap.
• 🟡 (13:20) packages/core/test/embedding-pool.test.ts defines warmPool(fakes) by calling embed(["bootstrap"], "document"), flushing, requiring exactly 1 fake worker, calling fakes[0].completeAll(), and awaiting the warmup.
• 🟡 (13:20) EmbeddingPool dispatch (#999) test setup saves VOYAGE_API_KEY, OPENAI_API_KEY, LORE_EMBED_POOL_SIZE, NODE_ENV, and the current provider; deletes both API keys and LORE_EMBED_POOL_SIZE; calls _setConstrainedMemoryForTest(0) so real CI cgroup limits do not interfere with _setPoolFreememForTest; resets the local-provider probe; and clears the provider.
• 🟡 (13:20) EmbeddingPool dispatch (#999) teardown resets _setTestWorkerFactory(null), _setEmbedPoolSizeForTest(null), _setEmbeddingWorkerWatchdogsForTest(null, null), _setLocalInitCooldownMsForTest(null), _setPoolFreememForTest(null), _setAvailableMemoryForTest(null), _setConstrainedMemoryForTest(null), and _setRecallEmbedsInFlightForTest(0); resets local-provider probing and provider-shutdown tracking; restores the provider and saved environment variables; and calls vi.useRealTimers().
• 🟡 (13:20) The test serializes cold bootstrap before enabling parallel dispatch sets pool size 2 and free memory to 64 * GB, installs fake workers, then starts embed(["alpha"], "query") and embed(["beta"], "query"). Before a real result establishes model readiness, it requires exactly 1 worker and only 1 dispatched embed ID, preventing sibling initializers from racing the shared model cache.
• 🟡 (13:20) In the same cold-bootstrap test, after the first fake worker completes alpha, beta is dispatched serially on that worker. Once bootstrap succeeds, concurrent embed(["gamma"], "query") and embed(["delta"], "query") calls grow the pool to exactly 2 workers, each with exactly 1 embed ID, demonstrating that parallel dispatch begins only after model readiness is proven.